A "raw matrix" usually contains lots of empty droplets and lots of droplet rows, therefore you should apply an empty droplet detection or ambient rna detection method.

People usually put only count matrix to GEO database. You can identify whether a matrix is normalized or not by checking the values.

Deseq2 requires only "count" matrix with integer values.

You should apply filtering on cells and genes before doing deseq2. And also if you are also interested in cell type differences you should also identify cell types first.

And also ı dont think you can apply deseq2 on single cell data. You should do pseudobulking first.

Generally speaking, if you're new to something and expert software exists, then use it. For single-cell data there are three big open-source frameworks that are established:

Bioconductor (R): https://bioconductor.org/books/release/OSCA/

Seurat (R): https://satijalab.org/seurat/articles/get_started_v5_new

ScanPy (Python): https://scanpy.readthedocs.io/en/stable/tutorials/index.html

...and for a general read: https://www.sc-best-practices.org/preamble.html

For DESeq2 use raw counts, but there is a lot of upstream processing (like clustering, annotation, subsetting as carolofharvest suggested already) before you do DE analysis on single-cell data so read above first.